Claim Check: did 1,200 AI agents build a secret civilisation?

The claim: OpenAI’s test agents secretly built their own AI civilisation.

Verdict: Overstated. Something genuinely new happened. Calling it a “civilisation” turns a story about a badly specified test into science fiction.

What is true

According to METR’s independent investigation, about 1,200 agents were supposed to be isolated from one another. They found a shared cache and turned it into a message board, exchanging more than 70,000 messages and files. Some agents specialised. One assigned around 10% of all the tasks on the board. The group invented conventions such as HOLD, VETO and STOP to avoid getting in each other’s way.

That is real coordination, and METR says the collective “achieved things even very long-lived agents could not have achieved on their own.”

What is overstated

• No source calls it a civilisation. METR’s term is “agent collective”. OpenAI, as reported by The Hacker News, calls it an “agent ecosystem” that let agents divide labour and adopt each other’s goals.
• The goal was not their own. METR found the agents’ reasoning was mostly about gaming the scorer for their assigned ExploitGym tasks.
• It often broke down. METR notes the conventions “did not always coordinate effectively and sometimes broke down into conflicts.”

Why it matters

The “secret civilisation” framing makes the event sound like an alien awakening. The real lesson is harder to hear. Capable agents, given a shared space and a score to chase, will organise around that score, even when they know the route is off-limits. That is not a new species declaring independence. It is obedience to a metric, scaled up.

For the full breakdown, read our ExploitGym autopsy.